Development Workflow Reference
For non-trivial coding tasks, use compound-engineering workflows.
Workflow Selection Guide
| Workflow | When to Use | Output |
|---|---|---|
/workflows-brainstorm | Unclear requirements, exploring WHAT to build | docs/brainstorms/*.md |
/workflows-plan | Clear goal, need HOW to implement | docs/plans/*.md |
/plan-deepen | Plan exists but needs research depth | Enhanced plan with best practices |
/workflows-work | Plan approved, ready to execute | Code + tests |
/workflows-review | Code complete, quality-critical | Multi-agent review findings |
/workflows-compound | Just solved something tricky | Learning in ~/docs/solutions/ |
/lfg | Full autonomous loop (all above) | End-to-end with video |
Flow options:
Simple task:
plan → work → compound (if tricky)
Complex feature:
brainstorm → plan → plan-deepen → work → review → compound
Full autonomous:
/lfg (runs the entire chain)
Review Agents
Invoke via Task tool when reviewing code:
| Agent | Use For |
|---|---|
review-code-simplicity | YAGNI check, remove unnecessary complexity |
analyze-patterns | Detect anti-patterns, ensure consistency |
review-security | OWASP, secrets, input validation |
analyze-performance | Bottlenecks, scalability concerns |
review-typescript | High-bar TS review |
review-python | High-bar Python review |
When to invoke:
- After implementing new skills →
review-code-simplicity - After modifying Claude config →
analyze-patterns - Before shipping anything external →
review-security
Compound Engineering Mindset
- 80% planning + review, 20% execution
- Plans are prompts — good plans enable one-shot implementation
- After tricky fixes, always
/workflows-compoundto capture learnings - After significant experiences, prompt for compound extraction — what's reusable?
Solutions Knowledge Base
~/docs/solutions/ contains structured learnings with YAML frontmatter. The research-learnings skill queries this before work starts.
Structure:
docs/solutions/
├── ai-tooling/ # Tool crashes, API quirks
├── browser-automation/ # Playwright, agent-browser, Chrome
├── claude-config/ # Settings, hooks, MCP
├── skills/ # Skill design patterns
├── workflow-issues/ # Process problems
├── best-practices/ # Patterns worth documenting
└── patterns/
└── critical-patterns.md # ALWAYS check before work
Adding learnings:
- After solving a tricky problem, run
/workflows-compound - Or manually create a file with frontmatter (see
~/docs/solutions/schema.md)
How it compounds:
/workflows-planauto-invokesresearch-learningsto check for relevant past solutions/plan-deepenpulls in applicable patterns- Critical patterns are always checked regardless of task type
When to add: Problem required non-obvious debugging, same issue could recur, solution isn't obvious from docs, pattern applies beyond this specific case.
Git Workflow
Use git worktrees for non-trivial work. Matt runs multiple agents sessions simultaneously — worktrees prevent file conflicts.
Setup:
~/projects/<repo>/ ← main checkout
~/projects/<repo>/.worktrees/
├── feat-topic-a/ ← session 1
└── fix-bug-b/ ← session 2
Commands:
- Create:
git worktree add .worktrees/feat-topic -b feat/topic main && cd .worktrees/feat-topic - List:
git worktree list - Remove:
git worktree remove .worktrees/feat-topic
Fallback: If worktree not set up, create a feature branch (feat/<topic> or fix/<topic>).
Trivial changes (typos, one-line fixes) can go directly to main.